Camera–LiDAR: bounded nearest-image-edge distance

This is a post hoc development ablation on the same fourteen reacquired spent KITTI inputs. It tests whether bounded distance to image ridges reduces the incorrect matches supported by the original intensity-and-spill objective. It is not a new SOTA audit, and does not change estimator defaults or thresholds.

Fixed method

Use the original image downsampling, Gaussian smoothing (sigma 1 pixel), Sobel gradients and 99th-percentile gradient normalization. Retain gradient ridges by nonmaximum suppression along the local gradient and a fixed normalized magnitude threshold of 0.2. Exclude the image border. Compute the Euclidean distance to the nearest retained ridge and use exp(-0.5 * (distance / 2)^2) as the reward, setting it to zero beyond 6 downscaled pixels. These constants were fixed before running the experiment, without selecting against reference errors or held scores.

This replaces the complete image response, including its background subtraction; it is one objective-family ablation, not evidence isolating the contribution of each ingredient. Bilinear interpolation samples the distance-derived reward field. There is no orientation penalty, additional baseline restart, or match-count normalization. All original point weights remain in the denominator, even for points projecting outside the image or beyond the distance limit.

Nearest-ridge matching can still be many-to-one, can match texture rather than physical depth boundaries, and does not enforce visibility or temporal identity. The 6-pixel cutoff is on field samples; bilinear interpolation smooths its cell boundary. A narrow field may also reduce the convergence basin. These limitations are reasons to inspect fitted poses as well as detection counts.

Source a19f1a00 was frozen in an isolated worktree. Both arms retain training-only initialization, multi-start nuisance refitting and all original search settings. Four worker processes each run isolated input pairs. Every original arm must reproduce the previous training-only results within absolute 1e-9 before its changed arm starts. All fourteen conditions and any failures are retained.

OPENBLAS_NUM_THREADS=1 PYTHONPATH=src:. python -m tools.kitti_camera_lidar_boundary_batch \
  /path/to/capture --output /path/to/new/output --nearest-edge

Forty-two correlated controls from two drives are not forty-two independent observations. Training-fold disagreement is measured against dataset reference calibration, not ground truth. Worker timings include concurrency effects and are not a repeated performance benchmark.

Results and decision

Do not promote. All fourteen original-arm parity checks pass within absolute 1e-9 and all paired inputs complete. Detection coverage decreases without recovering any controls on 0002.

Diagnostic	Original scalar	Nearest ridge
Detected controls, 0005	19/21	13/21
Detected controls, 0002	0/21	0/21
Total detected controls	19/42	13/42
Worst active-axis fold reference disagreement among detected controls	0.350 deg	0.246 deg
Worst disagreement on the same 11 controls detected by both arms	0.350 deg	0.200 deg
Target 0002 pitch +3 control t	0.77397	-1.76402
Target pitch +3 detected	No	No
Sum of profile wall times (concurrent workers)	177.9 s	183.7 s

Eight previously detected controls are lost and two are gained. The smaller maximum reference disagreement therefore does not establish better overall accuracy or a superior detector. The shared eleven controls are also a selected subset, not an independent calibration evaluation.

For the targeted pitch input, all six fitted active-axis reference disagreements are within 0.052 deg, yet only two held-out folds beat both nuisance-refitted controls. Its mean objective drop is negative (-0.03365). A small fitted pitch error can coexist with incorrect poses preferred by the held-out image objective. The full-pose diagnostic subsequently shows baseline rotation norms of 0.638–4.717 deg: small pitch error does not establish a nearly correct full rotation. This is direct evidence that a bounded nearest-image-edge score does not by itself resolve the ambiguity on this input; it does not establish which physical surfaces caused it.

Keep the default scalar objective. Before another objective variant, inspect which image structures the wrong poses match and whether those correspondences persist across frames. Do not tune the distance cutoff or ridge threshold on these results as if it were independent validation.

The complete report, summary and hash manifest retain every input, arm, fold, image-ridge count, source digest and changed detection flag.